From stochastic systems to governed intelligence: a neurosymbolic path for pharma

Artificial intelligence is being deployed across the pharmaceutical value chain, yet returns have been inconsistent. One of the explanations for this inconsistency is the organizational focus on models rather than the data foundations needed to support reliable use.

This Critical Perspective argues that pharma's AI challenge is not primarily model capability, but rather representation and governance. The authors examine why model-centric deployment strategies underdeliver in semantically weak and operationally complex setting. The authors follow this by outlining a semantic reference architecture in which heterogenous enterprise data are organized around explicit semantics while existing operational systems remain largely intact. Within this framework, ontologies function as computational substrates for conceptual clarity, interoperability, and machine reasoning. On this basis, the authors position neurosymbolic AI as a practical deployment model in which neural services provide flexible interpretation and generation, while symbolic substrates ensure semantic coherence with a knowledge base, therefore governing what may be accepted into real workflows.

Durable value from AI in pharma will not come from layering increasingly capable generative models onto poorly governed data environments, nor from relying on human review alone to compensate for weak system design. It will come from operationalizing semantics as infrastructure so that neural outputs are inspectable, verifiable, and accountable before they affect real work.